Google's Gemini Tool Retrieval Preview Adds 26 Tokens Per Idle Tool and Publishes No Savings Figure
A.I. / news
Google's Gemini Tool Retrieval Preview Adds 26 Tokens Per Idle Tool and Publishes No Savings Figure
Anthropic and OpenAI each reported token reductions for the same defer_loading flag. Google's November 19 post reports none.

Google put Tool Retrieval into preview in the Gemini API on November 19, 2026, letting an agent hold back the parameter schemas of its tools until the model asks for one. The announcement publishes no token-savings figure, no accuracy result and no latency number.
The post is by Lucia Loher, a product manager for the Gemini API, and Nora Kassner, a research scientist at Google DeepMind. The feature works only in the Interactions API, not the older generateContent method, and requires Gemini 3.6 or later, including all Gemini 4 models.
How it works
A developer adds a {"type": "tool_search"} tool to the request and sets "defer_loading": true on the tools to hold back. The model then sees only each deferred tool's name and description. When it wants one, it asks by name, the API loads the schema, and the model emits the function call, all inside one request.
Google's post says "There's no index to build or host" for the server-side mode. The response steps show tool_search_call, tool_search_result and function_call, which the post says makes tool names and descriptions easier to debug.
Setting defer_loading: true on a remote MCP server defers every tool that server exposes. Developers with their own registry, embedding index or per-user permission filter can instead declare tool_search with "execution": "client". Each client-side search adds a round trip.

What the 26-token figure does and does not say
Google's documentation says each deferred tool adds about 26 tokens to the prompt "in early testing", according to PPC Land. Google gives no baseline for an average full schema, so the saving cannot be calculated from its material.
PPC Land's own arithmetic is that 100 deferred tools add roughly 2,600 input tokens per request, about $0.004 at $1.50 per million input tokens for Gemini 4 Hydrogen. That excludes loaded schemas and search passes.
Pricing and the costs Google lists
There is no separate charge. Tokens are billed at the standard rates of the model called. Enabling tool_search adds a small fixed block of input tokens, and every schema the model loads is billed as input for the rest of the conversation.
The post says savings occur when unloaded schemas outweigh loaded ones, "the usual case for large libraries". It adds: "Measure it against your own traffic."
PPC Land lists further limits from the documentation:
- Loaded schemas cannot be unloaded.
- Each server-side search adds a billed model pass.
- The full tool list must be resent on every request, even with
previous_interaction_id. - Server-side and client-side retrieval cannot be mixed in one request, and only one
tool_searchis allowed. - A failed match returns text rather than a structured error.
Anthropic and OpenAI shipped first
PPC Land's comparison shows all three vendors use the same defer_loading: true flag. The two earlier vendors published results from their own tests on their own models, so the figures are not comparable with each other.
| Vendor | Launch | Published token result |
|---|---|---|
| Anthropic | November 24, 2025 | 85% reduction, about 77,000 to about 8,700 tokens |
| OpenAI | March 5, 2026 | 47% reduction across 250 tasks on 36 MCP servers |
| November 19, 2026 | None |
PPC Land also reports Anthropic's accuracy results: Opus 4 rose from 49% to 74%, and Opus 4.5 from 79.5% to 88.1%. It found no OpenAI accuracy figure and Google published none.
Google's documentation lists four supported models and says others return a 400 error, PPC Land reported. It omits Gemini 4 Carbon and Gemini 3.7 Flash. The feature is in preview on the v1beta API, and the post gives no general-availability date. Other agent tooling we have covered includes Chrome's Mirror Dimension for testing web agents and OpenAI's Codex sandbox.
Sources
More in A.I.
- 01Thinking Machines Ships Inkling 2 Mini, a 26.4B Apache 2.0 Model Sized for One DGX SparkThe model card lists 3.4B active parameters, a 1M-token window and a 20GB checkpoint, but does not say who ran most of its benchmark rows.
- 02Anthropic Closed Its Debut at $97.17, 17% Over the $83 IPO Price, as Nasdaq Fell 2%The $81.34 billion offering opened 34% higher, then gave back most of the gain while Nvidia, AMD and Micron dropped.
- 03OpenAI Says New Safeguards Cut Reasoning Extraction From 54.0% to 0.2%The figures are from OpenAI's own red-team evaluation, and the post names DeepSeek and Moonshot AI as distillers of U.S. models.
- 04Frostbite, an AI-Designed CPU, Claims 4.59x Efficiency Over an Unnamed BaselineArchitect Labs reports pre-silicon results from a 7 nm layout and has not said the chip has been fabricated.